interpretations (Maltz, [1984] 2001). One of Maltz’s ([1984] 2001) concerns was that recidivism has been defined in ad hoc ways that do not fully consider the underlying meaning of what is being measured. More recent commentary points out how distinctions in definitions affect conclusions about the performance of corrections systems (Weisberg, 2014) suggests the need for uniformity in measurement to allow for comparisons of outcomes across jurisdictions (Chen and Meyer, 2020; Council of State Governments Justice Center, 2014), and at the same time demonstrates the utility of different measures and definitions (Rhodes et al., 2014).

This chapter explores the ways recidivism is calculated and reported, discusses the strengths and weaknesses of different approaches, and points to conclusions that can—and those that cannot—be drawn from reported recidivism estimates. Consistent with the scope of this study, the chapter focuses primarily on recidivism following release from prison, although recidivism following other criminal legal system contacts is mentioned. Recommendations based on the chapter’s findings and conclusions are presented in Chapter 5.

Careful review of the strengths and limitations of current measures of recidivism is important, because ignorance about how data are captured can lead to misuses in policy and practice. Different methodologies and sampling techniques, as discussed below, are needed to answer different kinds of questions related to offending behaviors and involvement with the legal system, and reliance on inappropriate samples can lead to erroneous conclusions. For example, recidivism rates that measure events (such as counting each case of admission to prison, in a particular window of time for which one individual could account for more than one event) provide different information from rates that measure populations (such as tracking post-release behavior of all those incarcerated during a particular window of time). This distinction between populations and events of interest is often lost, not only when reporting recidivism-related statistics but even when generating such statistics. Too often, errors can be made by those interpreting and relying on recidivism data to make policy and programmatic decisions within the criminal legal system.

ANNUAL PRISON RELEASES

We begin by exploring patterns of prison releases in the United States, one population for whom recidivism rates are regularly calculated, and then review Bureau of Justice Statistics (BJS) statistics on recidivism patterns for release cohorts. In the most recent year (2020) for which data were available from the BJS at the time this report was being written, just under 550,000 sentenced individuals were released from state or federal prisons. As seen in Figure 2-1, although this number is down from its 2008 peak of

more than 735,000 individuals, it remains four times as high as when the BJS began systematic recordkeeping in 1978, a year when 140,000 individuals returned to the community. The number of people incarcerated in state or federal prison has declined over the last 10 years as annual releases generally exceeded annual admissions. A notable exception to the dominant pattern of alignment between releases and admissions is clearly visible in 2020. In the midst of the COVID-19 pandemic, the volume of releases declined but kept pace with the prior trend, whereas the decline in the volume of admissions was larger than expected by an order of magnitude.

Prison recidivism rates are often measured by the proportion of individuals who left prison in a given year who are later rearrested or reincarcerated (e.g., see “Recidivism in Bureau of Justice Statistics Reports” below). The National Prisoner Statistics data used in Figure 2-1 provide measures of readmission to prison of persons on parole. Such reincarcerations may result from the commission of new crimes or from violation of supervised-release conditions. Among individuals who were under some form of post-custody community supervision (such as parole) and returned to prison custody, the number of those recorded as readmitted for violations of parole conditions has grown steadily over time relative to the number of those

recorded as readmitted for new crime commission (Figure 2-2). Over the period 1978–2020, growth in the share of prison admissions from parole is accounted for almost entirely by events recorded as parole technical violations rather than new crimes while on parole.

These patterns of readmission illustrate a type of measurement error that arises in using aggregate statistics from administrative data to characterize recidivism. Prison admissions arising from technical violations of conditions of supervision may result from failure to meet conditions imposed on persons supervised in the community following release from prison (such as drug test failures, failure to show up for meetings, or failure to pay fees), or they may arise from new crimes that trigger technical violations. A new crime may trigger a technical violation because conditions of community supervision support an order returning a person to prison if a new crime was committed or a person was arrested, with the execution of the order constituting the technical violation.

The National Prisoner Statistics data are not sufficiently precise to distinguish pure technical violations from violations resulting from new crimes.

Consequently, the extent to which pure technical violations are counted as instances of recidivism or recommission of crime is unknown. Self-report data from the BJS survey of prison inmates offer some insight on this. For example, using data from BJS’s survey of individuals incarcerated in state prisons in 2004, Pfaff (2015) showed that 68.3 percent of persons in prison on a technical violation reported that they were later returned to prison on a technical violation following an arrest for a new crime. Pfaff’s estimates are for a prisoner stock, which may overstate the number of persons in prison under those conditions relative to the flow from persons admitted for violations. Grattet, Petersilia, and Lin’s (2008) analysis of technical violations in California sheds light on the flow issue. Looking at more than 265,000 technical violations occurring in 2003-04, they found that 35 percent consisted of noncriminal or technical violations. The remaining 65 percent consisted of behaviors alleged to have violated the California Penal Code, with more serious violations (e.g., robbery, rape, first-degree burglary) accounting for 10 percent of the code violations. Their work indicates that the behaviors underlying events recorded as technical violations were primarily new offenses and arrests, rather than violations of conditions of supervision (e.g., failure to report, positive drug tests, failure to notify change of address, and so forth).

Violation of the conditions of community supervision, which may include new criminal behavior as well as violations of conditions of supervision, accounted for about a third of prison admissions during the late 1990s through 2011; their share fell in 2011 following the U.S. Supreme Court decision in Brown v. Plata that required California to reduce its prison population.

Individuals who violate parole contribute less to the size of the prison population than to the total number of prison admissions.¹ This is because individuals who violate parole serve less time in prison (on average) than persons admitted on a new court commitment. For example, 14 percent of state or federal prison inmates in 2016 reported that they were on parole at the time of the event leading to their current imprisonment (Beatty and Snell, 2021). If pure technical violators serve less time than those admitted on technical violations stemming from new crimes, their contribution to the size of the prison population would be much smaller. Parole supervision issues and their impact on post-release outcomes are discussed in Chapter 3.

___________________

¹ This text was changed after release of the pre-publication version of the report to correct an error regarding the relative contribution of individuals who violate parole to the total prison population.

___________________

² Harrell and colleagues generally found that the results from the self-report and administrative data were largely comparable.

• Readjudication or an arrest proceeding to sanctioning in a court;
• Reconviction or a finding of guilt for a new crime;
• Reincarceration or a jail or prison sentence following conviction for a new crime; and
• Return to prison or any prison confinement for either a new crime or technical violation (see Durose, Cooper, and Snyder, 2014; Durose and Antenagenli, 2021).

Individuals released from prison after serving a sentence of at least one year are eligible to be included in a BJS release cohort and are identified from reports by state departments of corrections to the National Corrections Reporting Program (NCRP), another BJS product. After drawing a stratified random sample from all eligible individuals, the person-level corrections records are linked using fingerprint-based identification numbers to arrest and prosecution data from state criminal history repositories—“rap sheets” or records of arrest and prosecution—as well as from the Federal Bureau of Investigation (to track out-of-state recidivism). The criminal history data obtained by BJS comprises felonies and misdemeanors and includes information on arrest charges, court dispositions, sentences to incarceration, and custody status. The criminal history data that BJS uses in its studies are the same records that police officers use to determine a suspect’s current criminal justice status (e.g., on probation, parole, or bail); that judges use to make pretrial and sentencing decisions; and that corrections officials use to make inmate classification decisions (Durose, Cooper, and Snyder, 2014). Due to state-level differences in tracking modifications to arrest charges or court dispositions, only the originating charges and dispositions are recorded. Information on returns to prison obtained from later rounds of the NCRP is used to supplement incomplete and inconsistent record-keeping in state repositories. With appropriate weighting and survey adjustment, the BJS recidivism program yields generalizable (to the included states) estimates of recidivism using multiple definitions, with corresponding margins of error, for individuals released from state prison and still living within the reference window under study.

In its most recent recidivism study, BJS collected information on a stratified sample of 92,000 people, representing individuals released from prison in 34 states in 2012 who had served a sentence of one year or more (Durose and Antenangeli, 2021). The sample was representative of about 70 percent of individuals released from state prisons in 2012 but is not nationally representative. Some of the findings from this cohort study are summarized in Table 2-1 and described below.

Over one-third of individuals in the 2012 release cohort were arrested for a new crime within one year of their release, three-fifths within three years, and 71 percent within five years (Durose and Antenangeli, 2021).

A very large share of these rearrests result in a conviction, as by the fifth year after release, over half of released individuals are convicted of a new crime, typically by plea bargains. Returns to prison are also common, with nearly 46 percent of released individuals sent back to prison for either a technical violation or a new crime within five years. Demographic subgroup estimates indicate that males, members of certain minority groups (Black, American Indian, and Alaska Native), and individuals who are younger at the time of release tend to have a higher likelihood of recidivism, no matter whether the measure reflects interactions with police (rearrest), prosecutors and courts (reconviction), or return to prison.

Among those rearrested within five years of release, the most common offense of rearrest was a public order offense. Over half (54%) of persons released from prison in 2012 were rearrested for a public order offense and 49 percent were rearrested for an “other public order offense” (first column of Table 2-2). The other public order offense category is an undifferentiated category comprising conditional release violations (which include the aforementioned technical violations and arrests for new crimes reported as technical violations) along with lesser felonies and misdemeanors.³ The next most common arrest offense categories, each characterizing just over 20 percent of released individuals, are assault, larceny or motor vehicle theft, and drug possession.

Table 2-2 shows rearrest offenses by the most serious offense of commitment of those released from prison in 2012. The top row gives the total percent rearrested within each commitment offense category, and the subsequent rows give the percent rearrested by offenses of rearrest. The overall likelihood of rearrest is lowest for individuals who served a prison sentence for violent offenses (65%) and highest for property offenses (78%) (first row in Table 2-2). The single most common group of rearrest offenses is public order offenses, overall and irrespective of the nature of the commitment offense. Arrests for violent offenses accounted for 28.3 percent of all rearrest offenses. Those released from prison with violent offense charges were rearrested for a violent offense at slightly higher rates (32.4%) than those released on property (29.6%) or public order changes (28.1%).

To the extent there is evidence of crime specialization (or rearrest for an offense within the same class as their commitment offense), it is strongest for persons released from prison for property, drug, and public order offenses. This tendency is least pronounced for individuals released after serving a prison sentence for a violent offense, of whom about one-third

___________________

³ Durose and Antenangeli (2021, p. 25) define other public order offenses to include probation and parole violations, obstruction of justice, contempt of court, failure to appear, commercialized vice, nonviolent sex offenses, liquor law violations, bribery, invasion of privacy, disorderly conduct, contributing to the delinquency of a minor, and other miscellaneous or unspecified offenses.

are rearrested within five years for a new violent offense, followed by drug commitments, of which 43 percent are rearrested for a new drug offense. The BJS results on specialization are consistent with Pennsylvania data, which show modest degrees of specialization that are highest among property and drug offenses (Bell et al., 2013).

To summarize, while we do not have a national recidivism rate for individuals returning from prison, due to both coverage issues (the limited number of states providing data) and measurement issues (varying definitions, varying measures), the BJS release cohort recidivism program represents the best effort to provide that information. To date, BJS has standardized data collection for 34 states, representing 79 percent of all individuals released from state prisons in the United States (Durose and Antenangeli, 2021). Notably, states differ widely in what constitutes a punishable violation and how (or whether) that information is stored in criminal history repositories in the states that participate in the BJS cohort studies. Close inspection of rearrests indicates that while rearrests for violent crimes exceed the proportion of individuals convicted of violent offenses in the release cohort, many instances of recidivism result from other public order charges that do not necessarily align with measures of serious criminal behavior. These other charges include charges for violations of conditions of supervision, reflecting the operations of the criminal legal system and not necessarily having implications for public safety.

In all of its reports, BJS disaggregates rearrests by type of charge and reports on various characteristics of members of its release cohorts. Drawing on the criminal history records it obtains through the FBI’s Interstate Identification Index, BJS also reports on rearrests occurring outside the state in which an individual is released (Durose, Cooper, and Snyder, 2014; Durose, Snyder, and Cooper, 2015). This expands the scope of events covered beyond those included in studies using state-specific criminal history records. When BJS reports on reconviction and reincarceration rates, it generally reports on the cumulate rates of reconviction or imprisonment across the years of its follow-up periods. BJS reconviction statistics are limited to arrests resulting in reconviction.

Recidivism in Academic Literature

Many other studies also rely on official records of arrests, convictions, and imprisonment but vary in the specific measures and periods. Rearrest rates are commonly measured over one or two years and for up to eight years (Bird et al., 2021; Ford and Rector, 2020; Hunt and Dumville, 2016; Seigle et al., 2014). Rather than simply report an overall or summary recidivism number, most of these studies disaggregate rearrests by class or emphasize a particular class of events. For example, in their study of recidivism following

the California realignment, Bird and colleagues (2022) focused on felony rearrests, thereby limiting their recidivism measure to arrests for more serious offenses. Studies conducted by the U.S. Sentencing Commission on persons released from federal prison report rearrest rates for up to eight years and, like BJS, report the annual and cumulative rates, along with median times to first event and recidivism offenses (Hunt and Dumville, 2016; Cotter, 2021). Ford and Rector’s study (2020) of the Hawaii Opportunity Probation Evaluation (HOPE) measured rearrests over a one-year period, as did Seigle and colleagues (2014) in their study of recidivism following juvenile placement.

Reconviction measures may appear less frequently in recidivism studies, but generally when they appear, they accompany rearrest measures (e.g., Bird et al., 2022; Durose, Cooper, and Snyder, 2014; Hunt and Dumville, 2016). Recidivism rates that define recidivism as return to prison are commonly used in evaluations of the performance of corrections systems, with a three-year return-to-prison rate appearing commonly (Durose, Cooper, and Snyder, 2014; Gelb, 2018; Hunt and Dumville, 2016; King and Elderbroom, 2014; and Pew Center on the States, 2011, who also report prison return rates by year). A comprehensive recidivism study conducted by the Pew Public Safety Performance group analyzed data for three release cohorts (2005, 2010, and 2012) from 23 states and tracked returns to prison within the state of release for up to five years (Gelb and Velazquez, 2018). The Pew group reported reductions in recidivism rates, as measured by return to prison within three years of release, of nine and 13 percentage points from the 2005 base of 48 percent.

The extent to which studies measuring return-to-prison rates explicitly disaggregate by type of return varies—that is, whether they disaggregate among court commitment, a technical violation, or a new crime covered by a technical violation. State prison population forecaster measures of recidivism typically include readmissions for a parole violation (Harrison, 2021; O’Neil and Koushmaro, 2020; TenNapel et al., 2021) that may include both violations prompted by new crimes as well as technical violations (Hooks, nd). As Gaes and colleagues (2016) point out, distinguishing between a technical violation of a condition of supervision and a technical violation for reasons of a new crime is difficult with the data elements commonly found in corrections administrative databases. They accordingly recommend more extensive data. As noted earlier, evidence suggests that many if not a majority of events recorded as technical violations are actually new crimes or arrests of persons on parole, where the conditions of parole lead to a technical violation for a new crime (Grattet, Petersilia, and Lin, 2008).

The use of different measures of recidivism can cause confusion if one is looking to find out if recidivism rates have increased or decreased. Different studies use different recidivism events, different measures of the severity that constitutes recidivism, and different time periods over which

recidivism is measured. They also differ according to whether events occur during periods of correctional supervision or not. These inconsistencies require users of the research literature to take care in interpreting results. Additionally, each of the several measures of recidivism—rearrest, reconviction, or reincarceration for technical violations or new crimes—has limitations. Arrests that do not result in convictions can mean that “the usual suspects” were rounded up but none actually committed a crime. Misclassifying as pure technical violations new crimes that led to a violation can result in underestimates of the severity of behavior.

Although the varieties of definitions of recidivism present challenges for comparing outcomes across places or over time, the use of multiple measures of recidivism has utility. For example, Harding and colleagues (2017) studied the effects of imprisonment on recidivism measured both by reconviction and by reimprisonment. They found no impact of imprisonment on recidivism as measured by reconviction but found an impact on recidivism when measured by reimprisonment. They were able to attribute the difference to technical violations, such as failure to comply with parole restrictions, rather than new criminal behaviors, illustrating a type of analysis that can help illuminate the extent to which recidivism arises from the decisions made by criminal legal system actors versus new offense behaviors.

Comparing outcomes across samples and locations can contribute to an understanding of what may work to help reduce recidivism events. For example, in their review of evidence on the impacts of post-conviction imprisonment on recidivism Loeffler and Nagin (2022) include studies of rearrest and reconviction covering different follow-up periods, including rearrest for periods that include 18 months and 1, 2, 3, 4, 5, and 10 years; reconviction over a 5-year period; and reincarceration within 2 or 3 years. Their review examines differences in recidivism across correctional settings, such as those that give greater emphasis to rehabilitative programming, and they find that the settings that emphasize rehabilitative programming generally lead to less recidivism. While their study does not focus on the “recidivism rate,” it takes advantage of the fact that in the studies they reviewed, the authors explicitly defined their recidivism measures. Other cross-jurisdictional comparisons of recidivism have sought to measure the specific deterrent effect of incarceration on future offending (e.g., Nagin, Cullen, and Johnson, 2009; Roodman, 2017; Villettaz et al., 2006). All reached a similar conclusion, that incarceration has at best a null or mildly criminogenic effect on future offending, while Loeffler and Nagin’s (2022) review pointed to the importance of rehabilitative programs in achieving recidivism reduction effects. Despite differences in measurement, careful consideration of the definitions and measures used in cross-jurisdiction studies can allow one to draw conclusions about factors that affect recidivism.

post-release outcomes; here we address some measurement issues that recidivism studies need to consider. Among the most salient limitations of recidivism measures drawn from criminal legal system actions is that the measures reflect the interaction between a person’s offending behavior and the system actors’ responses to that behavior. We do not attempt to reconcile this discrepancy, but this aspect of measurement error is discussed in detail below.

Purposes and Uses

Central to the measurement of recidivism are the purposes to which measures are put. The purposes and uses determine the samples to be studied, the events to be measured, the durations between them, and the risk environment. Common purposes include program evaluation, program monitoring, performance measurement, forecasting of prison bedspace needs, and research about the correlates of recidivism. Different purposes may impose different requirements on measures and their interpretation. For example, in a drug court setting a treatment provider may want to measure substance use behaviors longitudinally to identify relapse and take appropriate responses. Or an evaluation of the effectiveness of in-prison programs that address criminogenic needs may study persons released from prison in different risk environments (e.g., measured by crime rates or level of police surveillance) to estimate future contacts with criminal legal system agencies. Studies that look at the performance of programs need to be clear about the follow-up periods. For example, if persons are released into parole, should the follow-up period be limited to the period of supervision or extend beyond it? These different periods not only have implications for measured recidivism rates, but they are linked to different research questions about supervision.

Samples and Populations of Interest: Event-Based and Person-Based Methods

The samples used to study recidivism need to be specified relative to the study purposes and derived from the populations about which inferences are to be made. If a study is interested in focusing on the outcomes of a group of people involved in some treatment program, such as an in-prison substance abuse program, the sample should represent the population of persons involved in that program. If a study seeks to understand the recidivism of persons who entered prison at any point in time, the sample should represent the entering cohort for the period under consideration, even if some members of the entering cohort might still be incarcerated at the end of the study period.

___________________

⁴ See: https://www.healthypeople.gov/2020/topics-objectives/topic/social-determinantshealth/interventions-resources/incarceration#:~:text=The%20U.S.%20releases%20over%207,people%20from%20prison%20each%20year.&text=However%2C%20recidivism%20is%20common.&text=Within%203%20years%20of%20their,than%2050%25%20are%20incarcerated%20again.

population of all persons incarcerated during the period. Under these conditions, a single exit-cohort-based estimate of recidivism will include a larger proportion of higher-recidivism-risk persons than the population of all persons incarcerated. This contributes to exit-cohort estimates of recidivism that may be higher than those obtained from an entering cohort or from the population of all persons incarcerated during a specific period (Rhodes et al., 2014).

As shown in Table 2-3 below, recidivism rates for event-based release cohorts are higher than those for person-based samples. The event-based releasees have a 21 percent return to prison rate in year 1 with variations across the 17 reporting states. The 2005 cohort has a slightly different reincarceration rate, with 24 states reporting. In both the event-based and person-based samples, most people released from prison do not return to prison within the observation window. In the 17-state event-based sample, about half did not return to prison. In the person-based sample, two-thirds did not return.

combination of behavior and justice agencies’ responses. Although rearrest, reconviction, and reincarceration are commonly used measures of recidivism events, each can be further subdivided by important attributes. For example, rearrest measured by specific charges helps in determining the severity of the arrested behavior. Theoretically, one could expand upon the concept of rearrest to measure the crimes cleared by an arrest or arrests that were not exceptionally cleared by prosecutors. Similarly, reincarceration rates can be refined to distinguish those following from a new sentence from those that did not and, if the data allow, the reasons for technical violations that result in return to prison.

These distinctions among categories of events that measure recidivism indicate that measures of events derived from criminal legal agency records reflect both the actions of criminal legal system officials and the behavior of individuals. Great care is required in making inferences from criminal legal contacts to identify a person’s offending behavior. Measuring recidivism in terms of new contacts with criminal legal system agencies is not necessarily equivalent to measuring recidivism as re-offending. Self-report data on offense behaviors may avoid this problem of criminal legal actors’ responses, but it can introduce measurement error and bias of its own in estimating the incidence or prevalence of reoffending. Regardless of the source of data, measurement error and potential bias in estimates are major methodological challenges confronting recidivism studies.

Frequency and Duration

Recidivism rates attempt to measure whether a set of individuals has engaged in further criminal behavior over a particular period of time. Reporting on recidivism rates often includes statistics such as time to a first event—the duration of time between release and an individual’s first recorded criminal behavior (their first “recidivist event”). Recidivism can also be tracked in terms of patterns or trajectories of recidivist events. Recidivism trajectories that decrease over a duration imply desistance from recidivism.

Connecting the length of a follow-up period to a program or policy purpose may not be simple or obvious. For example, recording recidivism events while a person is under supervision requires a different follow-up period than doing so after they leave supervision. Expecting a program to have long-run effects and therefore measuring recidivism over long periods of time may not reasonable; rather, a shorter follow-up period or sufficiently discrete periods to allow for an understanding of recidivism trends may be more appropriate. In some circles, the three-year follow-up rate has become an implicit standard, reflecting a trade-off between timeliness concerns and allowing a sufficient amount of time to pass for the slope of the

rate curve to flatten. For example, the Virginia Department of Corrections routinely reports three-year reincarceration rates (Virginia Department of Corrections Research—Evaluation Unit, 2021).

Ideally, the time frames for follow-up periods should be driven by the theoretical constructs or substantive aims of each recidivism study, but theory may not be sufficiently robust to suggest explicit lengths of followup periods; rather, theory may simply indicate that longer or shorter periods are appropriate. A reasonable approach might be for researchers to clearly state their aims, give a rationale for the length of follow-up periods, and maintain information on the timing of events so that the number of events up to different durations can be reported.

DATA SOURCES FOR MEASURING RECIDIVISM

The variations in how recidivism is measured also depend on the sources for the data. The two most common sources are self-report data and official records in administrative data maintained by criminal legal agencies. Table 2-4 summarizes the strengths and limitations of these measures. Private sources, such as consumer reporting agencies, may provide criminal background check information, but we exclude them from this review.⁵ The following section reviews the data sources for measuring recidivism and the adequacy of the measures.

Each source is generally associated with specific types of measures. Both classes of data sources have strengths and weaknesses. Self-reports may be a better measure of criminal behavior than administrative data but are costly to obtain and may suffer from recall biases. All classes of administrative data reflect the intersection of individual behaviors and criminal legal responses, and this source of measurement error may not be randomly distributed (see below for a detailed discussion). On the other hand, with administrative data larger samples are available, it is possible to identify trajectories, and the data have utility in demonstrating outcomes such as desistance (Blumstein and Nakamura, 2009) or periods of time when there are no legal system events.

Self-Report Data

Self-report data may provide several measures of recidivism events, primarily through self-reporting on criminal behavior. In addition, self-report data can be used to measure contacts with criminal legal agencies, such as arrests and convictions. Self-report measures of criminal behavior specific to drug use can also be used to evaluate the effectiveness of drug courts

___________________

⁵ For information about Consumer Reporting Agencies, see Lageson, 2020.

and treatment or aftercare programs to alleviate drug use and drug-related problems (Harrell, Cavanagh, and Roman, 1998); and to gauge drug use prior to incarceration in inmate surveys (e.g., Beatty and Snell, 2021).

Self-report data are typically derived from interviews with individuals or the collection of ecological movement assessment data (e.g., from smart-phones or emails). Persons are asked whether they engaged in certain offense behaviors. Questions about offense behaviors are typically based on descriptions of the events (e.g., types of crimes such as burglary, larceny, robbery) rather than legal codes.

Self-report data suffer from known problems related to precision and recall. Respondents’ reports of offending may be influenced by modes of survey administration, characteristics of the interviewer, anonymity, use of techniques to reduce response bias, and the length of the survey instrument (Gomes et al., 2019). Descriptive terms may have different meanings to different respondents. For example, one person may think of or characterize a burglary as a robbery even though the event was breaking and entering and did not involve use of force to take property. Self-report surveys that allow respondents to self-define criminal events or to affirm behaviors that fall into broad categories introduce measurement error into the classification of events. Alternatively, attribute-based interviewing uses cues to identify events, and the responses to cues result in an event’s classification into an offense category. For example, rather than ask a person if they committed a burglary, a respondent may be asked if they “broke into or attempted to break into a home by forcing a door or window, jimmying a lock, cutting a screen, or entering through an open door or window” along with questions about items stolen following entry.

Respondent recall problems have been studied at length, especially in national surveys such as the National Crime Victimization Survey (Bureau of Justice Statistics, 1989; Cantor et al., 2021; Rand and Catalano, 2007) and surveys of incarcerated individuals (Marquis and Ebner, 1981; Peterson et al., 1982). More prominent and rehearsed events are more likely to be recalled than less prominent events. Consequently, more serious offending behaviors are more likely to be recalled than, say, a rash of petty crimes such as shoplifting, larcenies, and simple assaults. If events are to be dated, two forms of telescoping may affect the dating of events. Forward telescoping includes events as having taken place during a specified time frame that was more recent than the event occurred. Backward telescoping includes events reported as occurring at a less recent time than specified.

The reliability of self-report data is improved by bounding interviews (as in the National Crime Victimization Survey) or by calendaring, by time line follow-back and anchoring procedures, and by the inclusion of

redundant questions as a way to check the reliability of responses (Peterson et al., 1982). If studies are designed to re-interview persons over time, self-report data are subject to attrition (loss of sample) and panel bias or respondent fatigue, which means that a respondent provides less information as the number of interviews increases.

Self-report surveys also have recruitment and nonresponse challenges. Sampled participants may choose not to participate for many reasons even if an honorarium is offered. Finding sampled participants to conduct an interview is challenging, especially for studies of recidivism, where sampled persons are generally known to be highly mobile. For program evaluations, conducting interviews at a program site, such as a drug court or probation office, may give an appearance of coercion whereby the sampled person thinks that she or he must participate in the survey as part of a program even if participation is voluntary. Despite researchers’ pledges of confidentiality to respondents, persons under supervision may be inclined to under-report criminal activity if they believe disclosing these activities could lead to violations. These challenges affect inferences about self-report offense behavior. The inferences can be improved if sources of nonresponse bias are accurately identified and addressed.

Inaccurate reporting of events is another concern with self-report data. Respondents may not report all criminal activity or may report some of it as less serious than it was. Alternatively, they may be predisposed to “boast” about behaviors and describe what they did as more serious than it was in reality. In their analysis of participants in the Cambridge Study in Delinquent Development, Auty, Farrington, and Coid (2015) found that respondents who had several convictions or convictions for more serious offenses were more likely to under-report them, while older persons were less likely to over-report seriousness. Overall, the authors found a high level of concurrent validity between the self-report and official records.

A final concern about inaccurate reporting relates to concerns that respondents may have in talking about ongoing criminal activities. This is especially the case if respondents report activities that could be suspected to be child abuse. Depending upon state law, researchers collecting these data may be mandatory reporters.

Despite the challenges of obtaining reliable self-report data, self-report data are of value, as this type of person-specific data can provide details not otherwise available about post-release behaviors, including the context or risk setting. But most recidivism studies do not use self-report data because of their cost and the need for skilled staff to conduct the interviews.

For example, if two individuals engage in the same criminal activity but the second lives in an area with a larger police presence, that second individual may be rearrested while the other’s crime goes undetected. Differences in the rate at which victims report crimes to the police affect the likelihood of an arrest. An individual may also be wrongfully rearrested or reconvicted of a crime they did not commit and still appear in recidivism rates.

of charges are usually available. The detailed charging information provides for the capacity to measure some aspects of the severity of arrests, although the analysis of charges within categories of crimes is often complicated. Rearrest as a measure of recidivism is commonly used in national-level recidivism studies (such as the BJS studies) and program evaluations.

The strength of rearrest measures lies in the official nature of the records representing local law enforcement agencies’ records of events. The records contain rich details about arrests, although such details are often not used in recidivism studies. The weaknesses of arrest data derive in part from the local-agency origination of the data. Jurisdictions vary in statutory classifications of criminal events, including designations of felony and misdemeanor statutes, which may present challenges in cross-site comparisons. State requirements on the non-felonious arrests to report to the criminal history repositories differ, so the scope of what is included in non-felonious arrest records also varies. Dispositions of arrests are incomplete. At the local level, arrest records may be linked to clearances of offenses, including exceptional clearances, but they do not include the disposition of all arrests. In the state repositories, final dispositions of arrests are missing in about 30 percent of all arrests; in the states with the highest disposition-reporting rate, this drops to 20 percent (Goggins and DeBacco, 2020). This means that arrest charges that are dropped either because a person did not commit the crime or the evidence was insufficient to move forward with a prosecution may be counted among rearrests.

State-level variation in what must be reported to repositories can add difficulties to making cross-state comparisons. And like all criminal legal system measures, arrests reflect a combination of a person’s behavior (e.g., criminal activity) and the response of law enforcement to that behavior. This cuts both ways, as some offenses do not result in arrests and some persons are arrested even though they have not committed a crime. Despite these weaknesses, efforts to reconcile self-report and arrest records as summarized by the National Research Council (2003) suggest that there is a high level of agreement between self-reports on having been arrested or having a police contact and having an official record. For more serious offenses and events, such as conviction, there is an even higher concordance between self-report and official records (Maxfield, Weller, and Widom, 2000; National Research Council, 2003).

Reconviction

Reconviction is a judicially determined event that occurs when an individual is found guilty of a criminal offense either by trial or by plea. Convictions are well measured in terms of the dates of event and the statutory charges and their dispositions. An administrative office aggregates data

from local courts in some states so that a single source can be accessed to obtain statewide data. This varies by state, and some states do not aggregate misdemeanor offenses. Court dispositions are reported to the states’ criminal history repositories, although complete disposition data may not exist in the repositories. Court records include information about each charge in a case when a defendant is charged with multiple offenses. This charge-specific information can be used in measuring the severity of the convicted offense. Some variations occur for how states handle misdemeanor offenses.

Reconviction measures exclude arrests that were not prosecuted. A conviction offense reflects the “bargained” or convicted offense behavior and not necessarily the behaviors that an individual engaged in. This bargained offense may be more or less serious than the underlying offense behavior. The determination of the convicted offenses reflects decisions of prosecutors, defense attorneys, and judges or juries, and the records of offense behaviors are based on criminal statutes, not offense-specific behaviors. Exceptions to this may exist under sentencing guidelines if the guidelines are based upon “real offense” behaviors, but in this case the real offense behaviors are applied at the sentencing and not the conviction stage. The distinction between the statutory classification of convictions and offense behaviors presents challenges in making inferences about offending behavior(s) when using conviction records. Reviews of self-report and official conviction records, however, have shown a high level of concurrent validity (Auty, Farrington, and Coid, 2015).

There are trade-offs in using reconviction and rearrest data in measuring recidivism. Using rearrests presents the risk of counting events in which a crime did not occur or that did not result in a conviction. Maltz ([1984] 2001) pointed to this concern by distinguishing any arrest from arrests that lead to conviction, and argued that when the latter is desired, it is incumbent on a user of administrative records to ensure that there is a conviction record for an arrest. Using arrests without disposition information presents a problem of false positives, because the measures then include as recidivism events acts that were not proven to be criminal.

Conversely, sole reliance on reconviction to measure recidivism presents the potential error of failing to capture data on an offense that did occur but for which charges were dropped or a conviction could not be obtained, for lack of evidence, witness cooperation, or prosecutor decisions not to move forward with a charge. This false-negative problem may understate the true level of criminal behavior. We do not know the extent to which the false-positive or false-negative errors present larger problems for recidivism estimates that are drawn from official records. Addressing this issue requires high-quality criminal history data (improvements to criminal history data are discussed below).

technical violations occur as a result of new crimes is not well understood, but the evidence suggests that most events recorded as technical violations include new crimes as the cause of the technical violation. We previously cited the study by Grattet, Petersilia, and Lin (2008) who evaluated technical violations in California and concluded that 65 percent involved behaviors alleged to violate the California Penal Code, and that about 10 percent of these were serious penal code violations. The BJS Annual Survey on Probation and Annual Survey on Parole reports that 6.5 percent of individuals exiting probation are incarcerated without a new sentence and 4.9 percent are incarcerated with a new sentence. The respective rates for those exiting parole are 11.2 percent and 5.3 percent (Oudekerk and Kaeble, 2021). The establishment surveys that BJS uses to obtain these data do not ask respondents to distinguish between individuals returned without a new sentence who committed a new crime versus those who committed a technical violation such as a drug test failure. Lattimore and colleagues (2016; 2018) report that technical violations are commonplace when individuals are being supervised, whereas arrests tend to be rarer.

As revocation for a technical violation requires either a judicial or executive (e.g., paroling authority) decision, records of dates of events and decisions are available. Information to link data on persons over time within jurisdictions is generally available, but in states where probation is organized at the county level, linkages of person-level records across places may be more challenging.

Consistent with other sources of administrative data, data definitions are not standard across supervising agencies, which increases difficulties in making comparisons across jurisdictions. From a practice perspective, supervising agencies and courts also differ with respect to their standards for revocable behaviors. Within jurisdictions, judges differ in deciding outcomes of probation revocation hearings. While variation across places and within jurisdictions has value for research purposes, for statistical purposes it presents challenges when the same event is treated differently by legal actors, agencies, and jurisdictions.

Graduated Sanctions

Graduated sanctions are also used by supervision agencies to manage noncompliance with requirements of probation or parole. Graduated sanctions involve increasing sanctions or requirements as a result of noncompliance with the supervision conditions of release. The completeness and accuracy of supervision agencies’ records of graduated sanctions are unknown. Local jails and state prisons will record entries into custody when ordered, but it is not clear that corrections departments’ data systems can distinguish between a commitment under a graduated sanctions

regime or another regime such as the imposition of a suspended custody sentence, technical violation, or new event. There is some question regarding whether the imposition of a graduated sanction is in fact a recidivism event, simply part of a sentencing package, or a management tool. Graduated sanctions policies are often not specified or uniformly applied (Rudes, 2012; Turner et al., 2012); consequently, their use reflects decisions by criminal legal system officials that may not systematically reflect the behaviors of the individuals on whom graduated sanctions were imposed. For example, Rudes (2012) found that parole officers resisted a rehabilitation-focused reform in California that discouraged the use of technical violations except in the most egregious cases through collaboration with police, the use of paperwork enhancement to encourage significant revocations, and by “piling” charges. Turner and colleagues (2012) similarly found that the implementation of a structured decisions-making tool for responding to violations of parole did not increase consistency in parole agent responses to violations.

Measurement Error

Earlier we described, in general terms, the sources of measurement error in both administrative and self-report data that measure recidivism events. All recidivism measures derived from administrative data reflect decisions by criminal legal system actors to take action and to record the actions taken in specific ways, as dictated by their roles in the criminal legal system and administrative records systems. As noted, these sources of measurement error present challenges for understanding the extent of recidivism events and in making comparisons across jurisdictions or over time. The issue is not whether administrative records will be used but how well they are used.

When recidivism measures derived from administrative data appear as dependent variables (e.g., what is the recidivism rate?), a general aim may be to make inferences about the true or underlying offending behavior. Measurement error can be additive or nonadditive. In the classic formulation of measurement error in a dependent variable, the measurement error is additive and is expressed as follows:

Y^* = Y + e,

where Y^* is the true, offense-based recidivism rate, Y is the observed arrest rate used to measure recidivism, and e is the random error, which is assumed to be normally distributed with zero mean and variance of one. Under these conditions, the expected value of Y^* equals the expected value of Y, and the arrest rate yields an unbiased estimate of the true recidivism rate.

We think it incumbent on the research community to examine the nature and effects of measurement error in administrative data. This may entail:

• conducting investigations of the processes that lead to the recording of events (such as arrests);
• designing studies that take into account differences among or within agencies in responses to offenses and how this affects the recording of recidivism events;
• further work on self-report and administrative data to better understand, in particular, how relatively minor offenses find their way into administrative data; and
• other research designs that improve our understanding of the effects of measurement error on recidivism estimates.

The sources and types of measurement error in recidivism measures require greater attention by researchers.

Efforts to Improve Administrative Data

We have shown that multiple and different measures of recidivism present challenges and opportunities for understanding. Used uncritically, multiple measures can cause confusion or misrepresent outcomes. Used critically, multiple measures can be analyzed and compared to generate conclusions about the impacts of incarceration on future offending and about the stages of the criminal legal system process responsible for outcomes. However, a challenge associated with multiple measures arises when the content of the data underlying a common term (e.g., percent rearrested) differ. Several efforts to improve administrative data focus on establishing common definitions of data elements.⁶ The value of commonly defined data elements across places and over time lies in facilitating making comparisons. We have argued that recidivism rates are commonly used to measure the performance of state corrections systems, and this naturally leads to questions of whether recidivism rates increase or decrease over time within states. If a state changes how it measures recidivism because new data are introduced, this presents problems for measuring change over time.

Analogously, states and other units of government are apt to compare themselves with other states. One reason they compare themselves is to learn if one entity is doing better and if so to find out why. While differences in what states measure may occur, when states use common outcomes but measure them differently, the comparison they desire will be unreliable. For example, if one state can distinguish between types of technical violations

___________________

⁶ See in this regard the Justice Counts initiative https://justicecounts.csgjusticecenter.org/.

and another cannot, but both report returns to prison for technical violations, the comparison will not be reliable. Hence, our primary concern is less with the measures and more with how the data elements used to create the measures are defined across places. We review several efforts to improve the measurement of violations of criminal law.

Efforts to develop national standards for criminal offenses as part of law enforcement statistics through the Uniform Crime Reporting (UCR) Program’s National Incident-Based Reporting System (NIBRS) are a step toward achieving some uniformity in the information local law enforcement agencies capture about crimes and arrests. NIBRS provides an incident-based data model for capturing detailed data on each crime incident and multiple attributes of arrests, victimization, and individuals involved in the criminal legal system. In 2019 (the latest year of data available at the time of the writing of this report), nearly 8,500 law enforcement agencies submitted NIBRS data to the FBI. The participating agencies were disproportionately smaller agencies. Collectively they represent 51 percent of the agencies that submitted data to the UCR program, but they covered less than 45 percent of the U.S. resident population. The reliability of NIBRS data across participating jurisdictions has yet to be fully assessed.

Efforts to develop standards for state courts include those promoted by the National Center for State Courts through its National Open Court Data Standards. Through the latter, the National Center for State Courts aims to develop business and technical court data standards to support the creation, sharing, and integration of court data by developing the rules by which data are described and recorded.⁷ Under this effort, states may still define events differently, but at a minimum the differences would be documented. The National Open Court Data Standards is not yet at an implementation stage, and the variability in record keeping continues to affect the use of court records in comparative studies of reconviction rates. To the committee’s knowledge, no such efforts are underway to document sentencing decisions.

Within states, the state criminal history record repositories collect and integrate records of arrests and prosecutions. These repositories provide informational services to the National Instant Criminal Background Check System, respond to requests for background checks on persons applying for jobs, and report data to sex offender registries, among other activities. While the repositories may provide a mechanism for achieving some uniformity in reporting arrests and prosecutions within states, the national criminal records exchange system faces well-documented shortcomings despite substantial investments by the federal government. The repositories’

___________________

⁷ See https://www.ncsc.org/services-and-experts/areas-of-expertise/court-statistics/nationalopen-court-data-standards-nods.

data integrate arrests with their prosecution and adjudication outcomes, but wide variation exists among states in the completeness of records that indicate the outcome of an arrest and in definitions of what records must be submitted to a state’s repository.

For example, some repositories obtain data on misdemeanors that others do not. The repositories have a program for enhancing data quality, known as the State Repository Records and Reporting Quality Assurance Program, which offers voluntary standards for information maintenance and reporting requirements. Promising efforts are currently underway by SEARCH and Rand to systematically assess data quality issues (Roberts, 2021), but they still have a long way to go.⁸ In the interim, the absence of uniform standards for criminal legal events results in measurement error, the full extent of which is unknown, that hinders comparisons within and across jurisdictions in the rearrest, reconviction, and reincarceration of persons released from prison.

RECIDIVISM AS BINARY: LIMITATIONS

Reporting recidivism in a binary way—sorting people into those who are and those who are not rearrested, reconvicted, or reincarcerated during the period of time being measured—gives an incomplete picture of a person’s post-release experiences. When a person’s re-engagement with the criminal legal system in any of these ways occurs, recidivism measures are interpreted to mean that person has failed, at least up to the point of the measured event. When it does not occur, the person has succeeded, as shown in studies of successful outcomes (Anderson, Schumacker, and Anderson, 1991; Peters et al., 2015). The same is true for criminal legal system programs: If the program reduces recidivism among its participants, it is considered promising or successful; if not, a naïve interpretation of the data implies that the program failed, whereas more sophisticated interpretations seek to find the reasons why a program that promised to reduce recidivism did not achieve the promised reductions.

When return to crime is measured simply by whether a person had a recidivism event or not, it limits efforts to understand post-release outcomes in the criminal legal system. An enhanced understanding of post-release outcomes occurs when recidivism studies address the seriousness, frequency, and trajectory of events (Lattimore, 2021). Breaking down recidivism rates by offense type (e.g., violent, property, public order) often represents an

___________________

⁸ Personal communication between committee member William Sabol and David Roberts, Executive Director of SEARCH, September 16, 2021. SEARCH is a national nonprofit organization of the States that provides resources for collecting, sharing, and analyzing justice information.

attempt to better capture seriousness, but there are also significant differences in seriousness even within offense categories. Reincarceration measures that do not distinguish between re-commitments for new convictions and those for technical violations of conditions of supervision conflate distinct behaviors. This conflation occurs even if technical violations are a signal that persons under supervision are failing to adjust their behaviors to community norms, portending a return to crime (Bushway and Apel, 2012), or if sanctioning technical violations is done to prevent more crime through incapacitation, specific deterrence, and general deterrence (Piehl and LoBuglio, 2005).

Dealing with the many events that are recorded in arrest and other records of criminal legal system actors is not easy. For example, NIBRS data report on 52 offenses in 23 categories. Criminal history records contain even more detailed offense information. Processing and reporting on this level of detail is not straightforward, and any classification system introduces heterogeneity within broader classes. Some guidance does exist for classifying offenses by severity, such as the attribute-based classification systems recommended by the Committee on National Statistics (Lauritsen and Cork (eds.), 2016). A major challenge associated with using attribute-based classification systems is that statutory charges do not capture all the elements of offenses. Nonetheless, attribute-based systems can provide guidance on thinking about measuring the severity of recidivism events.

Using official records to study recidivism trajectories has moved beyond the simple “yes/no” question of whether a person has a recidivism event. The redemption work of Blumstein and Nakamura (2009; 2010) illustrates this. Using official data to measure the arrest trajectories of individuals, Blumstein and Nakamura compared the trajectories of persons who had arrest records to the risk of arrest for same-aged people in the general population and to the risk of arrest for people who had never been arrested. Depending on the type of offense and age of first arrest, they found that the arrestee population had similar risks as the general population after 4.4 to 8.5 years. In other words, after a period of time the recidivism rates as measured by rearrest fell to the level of risk of arrest in the general population.

Similarly, the recent efforts by Bushway and colleagues (2022) on resetting risk illustrate the use of conviction records to demonstrate that most people with a conviction do not have a subsequent conviction. As the authors point out, current methods of measuring recidivism risk are based on the time of a person’s last conviction (or release from prison). Using this baseline, the measures do not adjust recidivism risk for the time a person has lived in the community without a new conviction. They find not only that most people with a conviction do not have a subsequent conviction, but that their risk of recidivism (measured by reconviction) declines considerably over the period from the last interaction with the criminal legal system.

7. Recidivism measures yield information on the presence or absence of negative outcomes and by themselves do not reflect the multivalent nature of post-release success, including employment, housing, health, family and community attachment, and personal well-being, which may be either associated with or independent of recidivism.

The existing literature on recidivism is a stepping stone to improve our measures of outcomes for persons released from prison. Our review signals precautions for future efforts to improve the data used to measure post-release outcomes and clearly and accurately communicate research findings. We offer the following guiding principles for future research and policy analysis:

1. The goals of the inquiry need to be clearly articulated, including how post-release success is linked to the research aims and how it is measured.
2. The samples used in studies need to be tied to the purpose of the studies. Studies of interventions or impacts of incarceration are good candidates for event-based samples. Studies that examine sentencing policy are good candidates for person-based samples. Studies of correctional performance that use event-based samples should consider how well the sample reflects the population of persons incarcerated who do not appear in an event sample taken in a given period.
3. Limiting analyses to simple, binary outcomes (whether someone did or did not engage in criminal behavior following release) without disaggregating by measures of severity or other salient correlates is an approach that should be avoided. Where data allow, time-dependent measures that track experiences of persons over time and allow for analyses of trajectories of behavior are preferred. Analyses of within-person outcomes over time place greater demands on the data, and efforts to create or facilitate access to these data are warranted.
4. The use of multiple measures of recidivism has utility for understanding how different recidivism events occur, both in studies conducted within a jurisdiction and in studies conducted between jurisdictions. The use of multiple carefully constructed and documented measures is warranted to help improve understanding of recidivism.
5. Improvements are warranted in administrative data and criminal history records to enable them to focus on distinguishing events, such as pure technical violations vs. new crimes or arrests that are reported as violations.

6. Generic recidivism measures that treat dissimilar behaviors or criminal legal system actions (e.g., felonies, misdemeanors, technical violations) the same way are best avoided. Researchers are best served by drawing on recent reports on modernizing crime statistics for ideas about taxonomies and classification of the many types of criminal legal system actions so that greater emphasis can be placed on offense-specific measures to assess the impact of a policy or program on criminal behavior.

As this chapter has demonstrated, there is a great deal of work to be done to improve the measurement of post-release criminal behavior. Addressing the limitations identified in this chapter will require coordinated effort between funders, researchers, and policy makers. Chapter 5 offers specific recommendations in this area, including for the development of more uniform standards for the measurement of success.

REFERENCES